AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Towards Distribution-Agnostic Generalized Category Discovery

Jianhong Bai, Zuozhu Liu, Hualiang Wang, Ruizhe Chen, Lianrui Mu, Xiaomeng Li, Joey Tianyi Zhou, YANG FENG, Jian Wu, Haoji Hu

NeurIPS 2023 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
regularize with true (oracle) long-tailed distribution instead of balance prior (ablation)BaCon pseudo-labeling branch · ImageNet-100-LToverall accuracy: 58.3 -> 64.6Improved
estimate distribution from contrastive branch (con k-means) instead of balance prior (ablation)BaCon · ImageNet-100-LToverall accuracy: 58.3 -> 62.0Improved
regularize only with labeled-set distribution instead of estimated distribution (ablation)BaCon · ImageNet-100-LToverall accuracy: 62.0 -> 44.8Hurt
k-means on classifier features instead of contrastive features for distribution estimation (ablation)BaCon · ImageNet-100-LToverall accuracy: 62.0 -> 56.4Hurt
add pseudo-label-based contrastive loss (vanilla) over baseline (ablation)BaCon · CIFAR-100-LToverall accuracy: 62.5 -> 65.5Improved
add debiasing of logits to vanilla pseudo-label loss (ablation)BaCon · CIFAR-100-LToverall accuracy: 65.5 -> 66.5Improved
add sampling of unlabeled instances to vanilla pseudo-label loss (ablation)BaCon · CIFAR-100-LToverall accuracy: 65.5 -> 66.6Improved
combine debiasing and sampling (ablation)BaCon · CIFAR-100-LToverall accuracy: 65.5 -> 67.2Improved
soft pseudo-label contrastive loss instead of hard loss (ablation)BaCon · CIFAR-100-LToverall accuracy: 64.9 -> 67.6Improved
dot product instead of L1/L2/cosine similarity for positiveness score (ablation)BaCon · CIFAR-100-LToverall accuracy: 62.7 -> 67.2Improved
vary re-estimate interval r from 1 to 50 (ablation)BaCon · CIFAR-100-LToverall accuracy: 67.9 -> 66.4No clear effect
vary hyper-parameter k from 0 to 1.5 (ablation)BaCon · CIFAR-100-LToverall accuracy: 66.2 -> 67.2Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy